We’re a small team of machine learning engineers working on personalized AI agents. As part of our research, we wanted to observe how humans edit documents generated by LLMs — which led us to build AgentCanvas, an open-source, canvas-like editor designed for capturing human editing traces and experimenting with alternative UX patterns beyond chat interfaces.
We explored tools like OpenCanvas, but found them too tightly coupled with LangChain and other heavy frameworks. So we built our own lightweight prototype — mostly hacking through TypeScript as ML folks
Highlights:
1. Built entirely on open-source libraries — e.g., blocknote.js for rich-text editing.
2. Uses simple chat completion APIs, no complex agent frameworks.
3. Designed to be easily modifiable for your own research or product ideas.
We’re looking for contributors! We’d love help from experienced FE/BE developers — especially those interested in the intersection of AI and UX.
Some open areas:
1. FE/BE separation – We're planning a Python backend to test agentic flows.
2. UX experiments – What kinds of UI (beyond chat) help users write/edit better? E.g., would memory blocks help with organizing context?
3. Better prompts, richer interactions, faster loading, you name it!
Repo: https://github.com/Mai0313/AgentCanvas
If this idea resonates with you, feel free to star the repo or share feedback. Contributions of all kinds are welcome